Singapore shares insights from testing of real-world generative AI apps

Singapore has unveiled insights from its Global AI Assurance Pilot, an initiative to catalyse emerging norms and best practices around technical testing of generative AI (gen AI) applications. These insights provided the blueprint for the world’s first Testing Starter Kit for gen AI applications. Both the Assurance Pilot and Starter Kit aim to uplift the capabilities of businesses in the safe deployment of gen AI applications and build overall trust in the AI ecosystem.

Source: IMDA. Senior Minister Tan delivering the opening keynote at the ATxSummit 2025.

The global initiatives were announced by Tan Kiat How, Singapore Senior Minister of State for Digital Development and Information at the ATxSummit 2025, the flagship event of Asia Tech x Singapore (ATxSG). Such efforts put Singapore at the forefront of efforts to operationalise AI safety, accelerate trusted and responsible AI adoption and deployment, and promote international cooperation for AI that benefits all.

"As a small, connected nation, we also want to contribute to a global AI ecosystem in support of open digital economies – one that is trusted, inclusive, and interoperable, building on our strong digital foundation and culture of trust. We are therefore building a home where AI can thrive, not just for ourselves but as a trusted node in a wider international community," Senior Minister Tan said in his opening keynote. 

"To realise this, we need to test what trust looks like in practice, so that it can be scaled with greater confidence."

The Global AI Assurance Pilot, an initiative by the AI Verify Foundation (AIVF) and Infocomm Media Development Authority (IMDA), was launched in February 2025 to encourage safe adoption of AI in industries and catalyse emerging norms and best practices around technical testing of gen AI applications. 

The pilot received strong interest from both local and international AI stakeholders, especially from companies deploying gen AI in their business processes. In the pilot, 16 specialist AI testers were paired with 17 deployers of real-world gen AI applications from 10 different industries including the finance, healthcare, HR, people and public sectors. 

A key finding from the pilot was that gen AI risks are often context-dependent (specific to industry, use case, culture, language and organisation). To narrow risks and tests for specific situations is a challenge; the recommendation is to involve subject matter experts throughout the application lifecycle. Senior Minister Tan shared more details:

"One, risks are context specific. Your risks depend on your use case. The most effective testing starts with a clear understanding of what is relevant to your use case, as well as what is not.

"Two, useful test data rarely comes ready-made. Most businesses do not have the prefect dataset sitting on the shelf. Generating realistic and edge case test scenarios requires thoughtful effort from us, with support from machines," he said.

"Three, go beyond outputs. Sometimes, the issue lies deeper in the pipeline. Testing what happens inside the system can offer more useful insights – and greater assurance.

"And four, large language models, or LLMs, can help with evaluation, but only with care. They can be fast and scalable, but still require thoughtful design, calibration, and human oversight. In some cases, simpler methods work just as well."

The results also showed that a human expert is essential at every stage of the testing lifecycle: from deciding what to test, to designing the right tests, to interpreting the test results and identifying improvement opportunities. The pilot has received a great deal of interest, with requests to run variations of the pilot and explore the introduction of technical standards. AIVF and IMDA will continue to work with sector regulators, international partners and industry players on next steps.

IMDA also announced plans to develop a first-of-its-kind Testing Starter Kit for gen AI applications. The Starter Kit generalises key insights from the Assurance Pilot and consultations with other practitioners to provide practical testing guidance for all businesses developing or leveraging gen AI applications, across sectors and use-cases. 

The Starter Kit provides a step-by-step guide on how to think about risks to be concerned about, highlighting common ones, and also how to test the gen AI applications. 

- Hallucination: the generation of incorrect content (factually inaccurate, lacks ground, incomplete)

- Undesirable content: generation of harmful content that inflicts harm on individuals, communities or the public interest

- Data disclosure: unintended leakage of sensitive information about individuals or organisations

- Vulnerability to adversarial prompts: susceptibility to producing unsafe output when presented with intentional prompt attacks

IMDA is calling for views from the industry on this Starter Kit on the testing guidance as well as recommended tests for the four identified risks.

The Starter Kit will continue to expand to address emerging risks and testing requirements in tandem with technological developments. Seven baseline tests from the Starter Kit have been made available on Project Moonshot,
enabling businesses to embrace responsible AI practices.

Singapore believes in harnessing AI for the public good and that AI can uplift economic potential, enhance social impact and meet the needs and challenges of our time. In line with this vision, AI Singapore (AISG) will be signing a memorandum of understanding (MoU) with the UNDP to advance AI literacy in six pilot countries aimed at closing the AI literacy divide and transforming communities in developing countries. This partnership will extend AISG’s successful AI for Good (AI4Good) programme – initially launched in 2024 to bolster national AI capabilities across Asia – to an international scale.

Asia Tech x Singapore (ATxSG) is Asia’s leading technology event jointly organised by IMDA and Informa. The event comprises three main segments, ATxSummit, ATxEnterprise and ATxInspire.

Details

Download the Starter Kit at https://www.imda.gov.sg/-/media/imda/files/about/emerging-tech-and-research/artificial-intelligence/large-language-model-starter-kit.pdf. 

The public consultation closes on 25 June 2025. Comments should be emailed to aigov@imda.gov.sg with the email header Comments on the draft Starter Kit for Safety Testing of LLM-Based Applications.